AI engineering
Retrieval systems that answer from your documents and show their working. An answer without a citation is an assertion, and in most of the places this gets deployed an assertion is not good enough.
We look at the documents before we look at the model. Format, freshness and who is allowed to see what decide the architecture, and they are the three things a demo never has to handle.
Chunking, hybrid retrieval and reranking are designed together against a question set your team wrote. Permissions are part of the query, never a filter applied to the answer afterwards.
Ingestion is a pipeline with re-runs and backfills rather than a script somebody ran once. Retrieval quality is visible in the interface while we are still building it.
Retrieval is measured separately from generation, because a wrong passage and an ungrounded answer are different bugs with different fixes. Refusing to answer counts as correct when nothing supports one.
Citations get clicked, and the clicks show where the corpus is thin. The evaluation set grows from real questions rather than the ones we imagined at the start.
Dense and lexical retrieval merged in one query, reranked, and raising rather than answering when nothing supports an answer.
"""Dense plus lexical, reranked, and never answered uncited."""
SEARCH = """
with dense as (
select id, 1 - (embedding <=> %(vec)s) as score
from passages where tenant_id = %(tenant)s
order by embedding <=> %(vec)s limit 50
), lexical as (
select id, ts_rank(tsv, plainto_tsquery(%(q)s)) as score
from passages where tenant_id = %(tenant)s limit 50
)
select id, sum(score) as score
from (select * from dense union all select * from lexical) hits
group by id order by score desc limit 20
"""
def retrieve(question, tenant):
vector = embed(question)
hits = db.fetch(SEARCH, q=question, vec=vector, tenant=tenant)
ranked = rerank(question, hits, model="rerank-v3")[:5]
# No passage, no answer. The citation is the contract, so
# refusing is a correct outcome rather than a failure.
if not ranked:
raise NoGroundedAnswer(question)
return rankedA retrieval assistant over clinical policy and procedure documents, answering staff questions with span-level citations and refusing to answer when the documents do not support one.
pgvector, in the Postgres you already run, unless the evals show you need more. One fewer system to operate is usually worth more than the last few points of recall — and if scale later justifies a dedicated store, the retrieval interface makes that a swap rather than a rebuild.
Retrieval-grounded generation with span-level citations, a faithfulness score in the eval suite, and an explicit path for the system to say it does not know. We treat an uncited claim as a bug, not a style preference.
Yes. Access is enforced at query time against your identity provider, so results are filtered per user rather than trusting a prompt not to mention something. Permission changes take effect on the next query.
Incremental re-indexing triggered by source changes rather than a nightly full rebuild, with freshness tracked as a monitored metric so a broken connector shows up as an alert instead of quietly stale answers.
A first call is 30 minutes and costs nothing. Bring the problem rather than a spec — the useful part is usually working out whether this is the right shape of solution at all.
contact@algologix.coWe reply within 24 hours.